The goal of our study was to assess the validity of the assumptions underlying three prominent workload models: the Time-Line Analysis and Prediction workload model (Parks & Boucek, 1989), the VACP workload model (Aldrich, Szabo, & Bierbaum, 1989), and the W/INDEX model (North & Riley, 1989). Sixteen subjects flew a low-fidelity flight simulation. Subjects were required to perform a two-axis tracking task, a concurrent visual-monitoring task, and a discrete decision task. The decision task had 16 variations defined by two levels on each of the following dimensions: input modality (visual vs. auditory), processing code (spatial vs. verbal), difficulty (easy vs. hard), and response modality (manual vs. voice). Dual-task costs were found only for the tracking task. The tracking data were then analyzed using two approaches: a traditional analysis of variance (ANOVA) and a correlational analysis of tracking performance versus model predictions. The ANOVA revealed that performance on the tracking task was better when the concurrent decision task was responded to vocally and was easy. Input modality and processing code of the concurrent decision task had no significant effect on tracking performance. The correlational analysis was used to evaluate each of the three models, to determine what features were responsible for improving the models' fit, and to compare their performance with a pure time-line model that makes no multiple-resource assumptions. All three models did a good job of predicting variance between experimental conditions, accounting for between 56% and 84% of the variance in our data and between 10% and 40% of an earlier data set. Different features of each model that affect the fit are then discussed. We conclude that it is important for models to retain a multiple-resource coding, although the best features of that coding remain to be determined. Coding tasks by their demand level appears to be less critical.